@ak--47/dungeon-master 1.5.3 → 1.6.0

package/dungeons/vertical/social/social.js

@@ -0,0 +1,1456 @@
+// ── IMPORTS ──
+import dayjs from "dayjs";
+import utc from "dayjs/plugin/utc.js";
+dayjs.extend(utc);
+import "dotenv/config";
+import * as u from "@ak--47/dungeon-master/utils";
+import * as v from "ak-tools";
+/** @typedef  {import("../../../types").Dungeon} Config */
+
+// ── OVERVIEW ──
+/*
+ * NAME:       Chirp
+ * APP:        Twitter+Instagram-style social media platform with algorithmic
+ *             feed, creator monetization, communities, and direct messaging.
+ *             Power users become "creators" with subscriber tiers; native ads
+ *             woven into feed and story placements.
+ * SCALE:      10,000 users, ~1.4M events, 121 days (2026-01-01 → 2026-05-01)
+ * CORE LOOP:  account created → profile updated → post created → consume + engage
+ *
+ * EVENTS (18):
+ *   post viewed (30) > post liked (18) > story viewed (15) > post created (12)
+ *   > notification received (12) > comment posted (10) > ad viewed (10)
+ *   > user followed (8) > dm sent (8) > search performed (7) > post shared (6)
+ *   > story created (5) > profile updated (3) > user unfollowed (2)
+ *   > ad clicked (2) > creator subscription started (2) > report submitted (1)
+ *   > account created (1)
+ *
+ * FUNNELS (8):
+ *   - Onboarding:       account created → profile updated → post created (70%)
+ *   - Feed Engagement:  post viewed → post liked → comment posted (45%)
+ *   - Content Cycle:    post created → post viewed → post liked → post shared (30%)
+ *   - Stories:          story created → story viewed → dm sent (40%)
+ *   - Discovery:        search performed → post viewed → user followed (35%)
+ *   - Notifications:    notification received → post viewed → post liked (50%)
+ *   - Creator Monetize: profile updated → creator subscription started → post created (15%)
+ *   - Ad Moderation:    ad viewed → ad clicked → report submitted (20%)
+ *
+ * USER PROPS:  app_version, account_type, follower_count, following_count,
+ *              bio_length, verified, content_niche
+ * SUPER PROPS: app_version, account_type
+ * SCD PROPS:   account_type (personal/creator/business/verified, monthly fuzzy, max 6),
+ *              community_status (new/growing/established/featured, monthly fixed,
+ *              max 6, community_id-scoped)
+ * GROUPS:      community_id (100 communities)
+ */
+
+// ── HOOK STORIES ──
+/*
+ * NOTE: All cohort effects are HIDDEN — discoverable only via behavioral cohorts
+ * (count an event per user, then measure downstream). No cohort flag is stamped
+ * on events. Algorithm-change source flips and engagement-bait short durations
+ * are raw mutations of config-defined props.
+ *
+ * -------------------------------------------------------------------------------------
+ * 1. VIRAL CONTENT CASCADE (everything)
+ * -------------------------------------------------------------------------------------
+ *
+ * PATTERN: Among users with 10+ pre-injection "post created" events, 5%
+ * become viral creators. Each of their posts triggers 60-120 cloned
+ * post-viewed, 60-120 post-liked, AND 60-120 post-shared events with
+ * offset timestamps (VIRAL_CLONES_MIN/MAX knobs).
+ *
+ * MEASUREMENT NOTE: the old "cohort A = 10+ post created" read is dead.
+ * At 1.2 events/user/day over 121 days the organic median is ~10 posts,
+ * and clone injections (H2/H6/H8) push ~85% of users past 10 final posts
+ * — binning on 10+ measures activity, not virality. The analyst-visible
+ * signal is CONCENTRATION: viral creators are ~3% of users but hold the
+ * bulk of all engagement events.
+ *
+ * HOW TO FIND IT IN MIXPANEL:
+ *
+ *   Report 1: Engagement Concentration
+ *   - Report type: Insights
+ *   - Events: "post viewed" + "post liked" + "post shared"
+ *   - Measure: Total per user, sorted descending
+ *   - Expected: top 3% of users hold ~55-68% of all engagement events
+ *     (measured 62% at reduced scale; organic counterfactual: 5.7%).
+ *     Viral creators show 500+ engagement events vs a ~45 median.
+ *
+ * REAL-WORLD ANALOGUE: A small share of creators drive a disproportionate
+ * fraction of all platform engagement.
+ *
+ * -------------------------------------------------------------------------------------
+ * 2. FOLLOW-BACK SNOWBALL (everything)
+ * -------------------------------------------------------------------------------------
+ *
+ * PATTERN: Users with 5+ "user followed" events have 50% of their posts
+ * duplicated with a 30-240 minute offset, plus an extra comment cloned.
+ * No flag — discover by binning users on user-followed count.
+ *
+ * HOW TO FIND IT IN MIXPANEL:
+ *
+ *   Report 1: Posts per User by Follow Activity
+ *   - Report type: Insights (with cohort)
+ *   - Cohort A: users with >= 5 "user followed"
+ *   - Cohort B: users with < 5 "user followed"
+ *   - Event: "post created"
+ *   - Measure: Total per user
+ *   - Expected: A ~ 1.5x posts per user vs B (measured 1.50 at reduced
+ *     scale)
+ *
+ *   MEASUREMENT NOTE: the 1.5x cohort ratio decomposes as ~1.24x activity
+ *   confound (high-follow users are simply more active — organic
+ *   counterfactual ratio) x ~1.21x true hook increment. The hook's +50%
+ *   applies to pre-injection posts only, so H6's +2-clones-per-post
+ *   stacking dilutes the incremental share for subscriber users. The
+ *   headline number still lands on ~1.5x because both factors compound.
+ *
+ * REAL-WORLD ANALOGUE: Users who actively follow many people tend to
+ * receive follow-backs and post more frequently.
+ *
+ * -------------------------------------------------------------------------------------
+ * 3. ALGORITHM CHANGE (everything)
+ * -------------------------------------------------------------------------------------
+ *
+ * PATTERN: On day 45 (2026-02-15), the dominant `source` for "post viewed"
+ * flips: before day 45, 70% of views are forced to "feed"; after, 70% are
+ * forced to "explore". Mutates an existing config-defined prop — no flag.
+ *
+ * INTERPLAY: H5 runs AFTER this flip and steals 30% of all post-day-30
+ * views for source="notification", so the post-day-45 explore share tops
+ * out near ~54%, not 70%. Three visible regimes (measured, reduced scale):
+ *   - pre-day-30:  feed 78%, notification 3%, explore 6%
+ *   - day 30-45:   feed 54%, notification 32%
+ *   - post-day-45: explore 54%, notification 32%, feed 5%
+ *
+ * HOW TO FIND IT IN MIXPANEL:
+ *
+ *   Report 1: Source Distribution Over Time
+ *   - Report type: Insights
+ *   - Event: "post viewed"
+ *   - Measure: Total
+ *   - Breakdown: "source"
+ *   - Line chart by day
+ *   - Expected: feed dominates pre-day-45 (~78% before day 30, ~54%
+ *     day 30-45); explore dominates post-day-45 (~54%); feed collapses
+ *     to ~5% after the flip
+ *
+ * REAL-WORLD ANALOGUE: Algorithmic feed redesigns shift content discovery
+ * from chronological to interest-based.
+ *
+ * -------------------------------------------------------------------------------------
+ * 4. ENGAGEMENT BAIT (event)
+ * -------------------------------------------------------------------------------------
+ *
+ * PATTERN: 20% of "post viewed" events get view_duration_sec collapsed
+ * to 1-5 seconds. No flag — analyst sees a bimodal duration distribution.
+ * Runs after all injection passes, so cloned views (H1/H6) are crushed at
+ * the same rate.
+ *
+ * HOW TO FIND IT IN MIXPANEL:
+ *
+ *   Report 1: View Duration Distribution
+ *   - Report type: Insights
+ *   - Event: "post viewed"
+ *   - Measure: Distribution of "view_duration_sec"
+ *   - Expected: ~ 20% of values cluster at 1-5 sec; rest at 5+ sec
+ *     (measured 20.4% at reduced scale; the organic generator floors
+ *     durations above 5s, so the low mode is 100% hook-made)
+ *
+ * REAL-WORLD ANALOGUE: Clickbait posts collect impressions but bounce
+ * quality is awful, dragging down avg watch time.
+ *
+ * -------------------------------------------------------------------------------------
+ * 5. NOTIFICATION RE-ENGAGEMENT (everything)
+ * -------------------------------------------------------------------------------------
+ *
+ * PATTERN: After day 30 (2026-01-31), 30% of "post viewed" events have
+ * source flipped to "notification". Runs after H3, so it wins on the
+ * overlap. Mutates the existing config-defined `source` prop.
+ *
+ * HOW TO FIND IT IN MIXPANEL:
+ *
+ *   Report 1: Notification-Sourced Views Over Time
+ *   - Report type: Insights
+ *   - Event: "post viewed"
+ *   - Measure: Total
+ *   - Filter: source = "notification"
+ *   - Line chart by day
+ *   - Expected: ~3% before day 30 (organic 20% share diluted by H3's
+ *     70% feed-force and by H1/H6 clones that never stamp notification),
+ *     then ~32% after (measured 3.3% -> 32.2% at reduced scale, ~10x)
+ *
+ * REAL-WORLD ANALOGUE: Push notifications about trending content are a
+ * primary lever for waking up dormant users.
+ *
+ * -------------------------------------------------------------------------------------
+ * 6. CREATOR MONETIZATION (everything)
+ * -------------------------------------------------------------------------------------
+ *
+ * PATTERN: Users with any "creator subscription started" event get 2 extra
+ * cloned posts and stories per original (3x rate), plus 25% extra cloned
+ * post-viewed events stamped `source="profile"`. No flag — discover via
+ * cohort builder.
+ *
+ * SCALE NOTE: "creator subscription started" is a weight-2 event over 121
+ * days, so ~78% of users have at least one — the cohort is an engagement
+ * marker, not an elite-creator flag. The 3x contrast still reads cleanly
+ * because the clone multiplier applies per-post. The `source="profile"`
+ * stamp is mostly overwritten by H3/H5 (which run later): only ~21% of
+ * the stamps survive, showing as a mild profile-share elevation
+ * (subscribers ~4.4% vs non ~2.3% of views), not a dominant source.
+ *
+ * HOW TO FIND IT IN MIXPANEL:
+ *
+ *   Report 1: Posts per User — Subscribers vs Not
+ *   - Report type: Insights (with cohort)
+ *   - Cohort A: users with >= 1 "creator subscription started"
+ *   - Cohort B: users with 0
+ *   - Event: "post created"
+ *   - Measure: Total per user
+ *   - Expected: A ~ 3.3x posts per user (3x knob compounded with a mild
+ *     activity confound; measured 3.32 at reduced scale)
+ *
+ *   Report 2: Stories per User — Subscribers vs Not
+ *   - Same cohorts, event "story created"
+ *   - Expected: A ~ 3.8x stories per user — a purer read than posts:
+ *     H2's snowball dupes lift non-subscriber POSTS (denominator) but
+ *     never stories, so the story contrast runs higher (measured 3.82)
+ *
+ * REAL-WORLD ANALOGUE: Creators with paying subscribers publish more often.
+ *
+ * -------------------------------------------------------------------------------------
+ * 7. TOXICITY CHURN (everything)
+ * -------------------------------------------------------------------------------------
+ *
+ * PATTERN: Users with 2+ "report submitted" events lose 60% of ALL
+ * activity after day 30 (2026-01-31) — each post-cutoff event dropped at
+ * 60% likelihood, including injected clones and the reports themselves.
+ * No flag — discover via retention or per-user activity drop.
+ *
+ * MEASUREMENT NOTE: compare each cohort's post-day-30 / pre-day-30 event
+ * ratio, not absolute counts — the dataset is growth-shaped (born-in
+ * users make post/pre ~3.2x for everyone), and reporters are heavy users
+ * whose absolute counts stay high even after the drop. Measured (reduced
+ * scale): reporters 1.27 post/pre vs normal 3.21 — a 0.40x relative
+ * contrast, matching the 60% drop knob. The organic counterfactual
+ * confirms the paired per-user drop is 0.42x exactly.
+ *
+ * HOW TO FIND IT IN MIXPANEL:
+ *
+ *   Report 1: Retention by Toxicity
+ *   - Report type: Retention (or Insights: post/pre day-30 event ratio)
+ *   - Cohort A: users with >= 2 "report submitted"
+ *   - Cohort B: rest
+ *   - Expected: A's post-day-30 activity ~ 0.4x of B's, relative to
+ *     each cohort's own pre-day-30 baseline
+ *
+ * REAL-WORLD ANALOGUE: Repeated reporters are signaling dissatisfaction
+ * and often quietly churn.
+ *
+ * -------------------------------------------------------------------------------------
+ * 8. WEEKEND CONTENT SURGE (everything)
+ * -------------------------------------------------------------------------------------
+ *
+ * PATTERN: 30% of Sat/Sun "post created" and "story created" events get a
+ * duplicate cloned 1-3 hours later. No flag — discover via day-of-week chart.
+ *
+ * BASELINE NOTE: the config's soup dayOfWeekWeights are all >= 1.0, which
+ * TimeSoup treats as accept-always (see lib/utils/utils.js TimeSoup phase
+ * 2) — the organic DOW distribution is FLAT (measured 0.98 weekend/weekday
+ * on the organic counterfactual). The 1.25x hooked ratio is therefore
+ * pure hook signal, slightly under the 1.3 knob because late-Sunday dupes
+ * (+1-3h) can roll into Monday and H7 drops some weekend clones.
+ *
+ * HOW TO FIND IT IN MIXPANEL:
+ *
+ *   Report 1: Posts by Day of Week
+ *   - Report type: Insights
+ *   - Events: "post created" + "story created"
+ *   - Measure: Total
+ *   - Breakdown: Day of week
+ *   - Expected: Sat/Sun ~ 1.25x weekday bars (measured 1.25 at reduced
+ *     scale)
+ *
+ * REAL-WORLD ANALOGUE: Weekend leisure time produces a natural creation surge.
+ *
+ * -------------------------------------------------------------------------------------
+ * 9. POST-CREATED MAGIC NUMBER (everything)
+ * -------------------------------------------------------------------------------------
+ *
+ * PATTERN: Users in the 3-7 pre-injection post-created sweet spot get
+ * +40% comment_length on their comment-posted events (richer engagement).
+ * Users with 8+ pre-injection posts get comment_length x0.7 (burnout —
+ * shorter, lazier comments). NO events are dropped; the effect is purely
+ * on the comment_length property. No flag — discover by binning users on
+ * post count.
+ *
+ * MEASUREMENT NOTE: the hook buckets on PRE-injection post counts;
+ * analysts see POST-injection counts inflated by H2/H6/H8 clones. With an
+ * organic median of ~10 posts, 73% of users are hook-"over" (burnout is
+ * the norm), and clone inflation pushes most boosted sweet-organic users
+ * into the 8+ final bucket, diluting it. The contrast survives because
+ * the final 3-7 bucket is dominated by boosted users who dodged clone
+ * inflation (non-subscribers). Paired-counterfactual per-user lifts are
+ * exact: low 1.00, sweet 1.40, over 0.68.
+ *
+ * HOW TO FIND IT IN MIXPANEL:
+ *
+ *   Report 1: Comment Length by Post Bucket
+ *   - Report type: Insights (with cohort)
+ *   - Cohort A: users with 3-7 "post created" (final counts)
+ *   - Cohort C: users with >= 8 "post created"
+ *   - Event: "comment posted"
+ *   - Measure: Average of "comment_length"
+ *   - Expected: A ~ 1.7x C (measured 250 vs 145 at reduced scale)
+ *
+ *   Report 2: Bucket Ordering
+ *   - Add Cohort B: users with 0-2 "post created" (small population —
+ *     unmutated users with almost no comments)
+ *   - Expected ordering: A (sweet, ~250) > B (low, ~200) > C (over, ~145)
+ *
+ * REAL-WORLD ANALOGUE: Moderate posters write thoughtful comments; the
+ * over-prolific burn out and leave short, low-effort replies.
+ *
+ * -------------------------------------------------------------------------------------
+ * 10. ONBOARDING TIME-TO-CONVERT (funnel-post)
+ * -------------------------------------------------------------------------------------
+ *
+ * PATTERN: Creator and business account_type users complete funnels
+ * 1.4x faster (gap factor 0.71); personal users 1.25x slower (factor
+ * 1.25). The hook intercepts EVERY funnel-post array (all 8 funnels, not
+ * just onboarding), computes the gap between consecutive steps, and
+ * scales each gap by the account-type factor before rewriting the step
+ * timestamps. The onboarding funnel (account created -> profile updated
+ * -> post created) is where the effect is cleanly measurable: "account
+ * created" is unique per user (isFirstEvent), so the funnel evaluation
+ * anchors on the scaled instance instead of racing organic standalones.
+ *
+ * MEASURED (reduced scale, emulateBreakdown timeToConvert, stable across
+ * 1h-24h windows): personal median ~23min, creator/business ~16min —
+ * fast/personal = 0.71. Per-leg attenuation vs the organic counterfactual
+ * (baseline ratio 0.96, medians ~20min): fast leg 0.71 knob -> 0.81
+ * realized, slow leg 1.25 -> 1.11; greedy step-matching occasionally
+ * substitutes organic standalone events for scaled ones, pulling both
+ * legs toward 1 (same mechanism as sass H9, milder here thanks to the
+ * unique first-step anchor).
+ *
+ * HOW TO FIND IT IN MIXPANEL:
+ *
+ *   Report 1: Onboarding Funnel Median TTC by Account Type
+ *   - Report type: Funnels
+ *   - Steps: "account created" -> "profile updated" -> "post created"
+ *   - Measure: Median time to convert
+ *   - Breakdown: "account_type" (superProp)
+ *   - Expected: creator/business median ~ 0.71x of personal
+ *
+ *   NOTE (funnel-post measurement): visible via Mixpanel funnel median
+ *   TTC. Cross-event MIN->MIN SQL queries on raw events do NOT show this.
+ *
+ * REAL-WORLD ANALOGUE: Creators and businesses arrive with clear intent
+ * and complete profile setup faster; personal users browse casually and
+ * take longer to commit to their first post.
+ *
+ * =====================================================================================
+ * EXPECTED METRICS SUMMARY (measured at reduced scale, 2K users; see stories
+ * export below for the NAILED/STRONG bands asserted at full fidelity)
+ * =====================================================================================
+ *
+ * Hook                      | Analyst-visible metric            | Measured
+ * --------------------------|-----------------------------------|------------------
+ * H1 Viral Content Cascade  | top-3% share of engagement events | 0.62 (organic 0.06)
+ * H2 Follow-Back Snowball   | posts/user, 5+ follows vs rest    | 1.50x
+ * H3 Algorithm Change       | post-d45 explore share            | 0.54 (pre-d30 feed 0.78)
+ * H4 Engagement Bait        | share of views <= 5 sec           | 0.204 (organic 0.00)
+ * H5 Notification Re-engage | source=notification share         | 0.033 pre-d30 -> 0.322 post
+ * H6 Creator Monetization   | posts/user sub vs non             | 3.32x (stories 3.82x)
+ * H7 Toxicity Churn         | post/pre-d30 ratio, rep vs norm   | 1.27 vs 3.21 (0.40x)
+ * H8 Weekend Surge          | weekend/weekday daily creations   | 1.25x (organic 0.98)
+ * H9 Magic Number           | comment_length sweet vs over      | 1.72x (250 vs 145)
+ * H10 Onboarding TTC        | funnel median TTC fast/personal   | 0.71 (organic 0.96)
+ *
+ * MEASUREMENT CAVEATS:
+ * - Paired organic counterfactual (hook: r => r, same seed) gives exact
+ *   per-user baselines: H9 lifts land on the knobs (1.00/1.40/0.68), H7's
+ *   drop is 0.42x paired, and the engagement-concentration jump (0.06 ->
+ *   0.62) is entirely injection-made.
+ * - Post-count cohorts use FINAL counts (analyst view). Clone injections
+ *   inflate these: 85% of users clear 10 final posts, 73% are hook-"over".
+ * - H10 is invisible to cross-event MIN->MIN SQL; use funnel median TTC
+ *   (emulateBreakdown timeToConvert) as in the stories export.
+ * =====================================================================================
+ */
+
+// ── SCALE ──
+const SEED = "harness-social";
+const NUM_USERS = 10_000;
+const DATASET_START = "2026-01-01T00:00:00Z";
+const DATASET_END = "2026-05-01T23:59:59Z";
+const EVENTS_PER_DAY = 1.2;
+const token = process.env.MP_TOKEN || "your-mixpanel-token";
+
+const chance = u.initChance(SEED);
+
+// ── KNOBS (tweak these to reshape stories) ──
+const VIRAL_POST_THRESHOLD = 10;
+const VIRAL_USER_LIKELIHOOD = 5;
+const VIRAL_CLONES_MIN = 60;
+const VIRAL_CLONES_MAX = 120;
+
+const FOLLOW_SNOWBALL_THRESHOLD = 5;
+const FOLLOW_SNOWBALL_LIKELIHOOD = 50;
+
+const ALGORITHM_CHANGE_DAY = 45;
+const ALGORITHM_FLIP_LIKELIHOOD = 70;
+
+const ENGAGEMENT_BAIT_LIKELIHOOD = 20;
+
+const REENGAGEMENT_START_DAY = 30;
+const REENGAGEMENT_LIKELIHOOD = 30;
+
+const CREATOR_POST_CLONES = 2;
+const CREATOR_STORY_CLONES = 2;
+const CREATOR_VIEW_CLONE_LIKELIHOOD = 25;
+
+const TOXICITY_THRESHOLD = 2;
+const TOXICITY_CUTOFF_DAYS = 30;
+const TOXICITY_DROP_LIKELIHOOD = 60;
+
+const WEEKEND_CLONE_LIKELIHOOD = 30;
+
+const POST_SWEET_MIN = 3;
+const POST_SWEET_MAX = 7;
+const POST_OVER_THRESHOLD = 8;
+const POST_SWEET_COMMENT_BOOST = 1.4;
+const POST_OVER_COMMENT_FACTOR = 0.7;
+
+const ONBOARDING_TTC_FAST = 0.71;
+const ONBOARDING_TTC_SLOW = 1.25;
+
+// ── DATA ARRAYS ──
+// Generate consistent post IDs for lookup tables
+const postIds = v.range(1, 1001).map(n => `post_${v.uid(8)}`);
+
+// ── HELPER FUNCTIONS ──
+function handleFunnelPostHooks(record, meta) {
+	// H10: ONBOARDING TIME-TO-CONVERT — creator/business 0.71x faster,
+	// personal 1.25x slower. Scales inter-step gaps via timestamp rewrite.
+	const segment = meta?.profile?.account_type;
+	if (Array.isArray(record) && record.length > 1) {
+		const factor = (
+			segment === "creator" || segment === "business" ? ONBOARDING_TTC_FAST :
+			segment === "personal" ? ONBOARDING_TTC_SLOW :
+			1.0
+		);
+		if (factor !== 1.0) {
+			for (let i = 1; i < record.length; i++) {
+				const prev = dayjs(record[i - 1].time);
+				const newGap = Math.round(dayjs(record[i].time).diff(prev) * factor);
+				record[i].time = prev.add(newGap, "milliseconds").toISOString();
+			}
+		}
+	}
+	return record;
+}
+
+function handleEverythingHooks(record, meta) {
+	const datasetStart = dayjs.unix(meta.datasetStart);
+	const userEvents = record;
+	if (!userEvents || userEvents.length === 0) return record;
+
+	// Stamp superProps from profile for consistency
+	const profile = meta.profile;
+	userEvents.forEach(e => {
+		e.app_version = profile.app_version;
+		e.account_type = profile.account_type;
+	});
+
+	// First pass: identify behavioral patterns (no flags written)
+	let postCreatedCount = 0;
+	let followReceivedCount = 0;
+	let reportSubmittedCount = 0;
+	let hasCreatorSubscription = false;
+	let isViralCreator = false;
+
+	userEvents.forEach((event) => {
+		if (event.event === "post created") postCreatedCount++;
+		if (event.event === "user followed") followReceivedCount++;
+		if (event.event === "report submitted") reportSubmittedCount++;
+		if (event.event === "creator subscription started") hasCreatorSubscription = true;
+	});
+
+	if (postCreatedCount >= VIRAL_POST_THRESHOLD && chance.bool({ likelihood: VIRAL_USER_LIKELIHOOD })) {
+		isViralCreator = true;
+	}
+
+	// Second pass: inject cloned events (no behavioral cohort flags)
+	for (let idx = userEvents.length - 1; idx >= 0; idx--) {
+		const event = userEvents[idx];
+		const eventTime = dayjs(event.time);
+
+		// H1: VIRAL CONTENT CASCADE — clone 60-120 view/like/share per post.
+		// Discovery: bin users by post-created count, observe per-user view/like/share volume.
+		if (isViralCreator && event.event === "post created") {
+			const viralViews = chance.integer({ min: VIRAL_CLONES_MIN, max: VIRAL_CLONES_MAX });
+			const viralLikes = chance.integer({ min: VIRAL_CLONES_MIN, max: VIRAL_CLONES_MAX });
+			const viralShares = chance.integer({ min: VIRAL_CLONES_MIN, max: VIRAL_CLONES_MAX });
+			const injected = [];
+
+			// Clone only from a real same-type template. The old
+			// `...(viewTemplate || event)` fallback would bleed post-created
+			// props (character_count, has_media, hashtag_count) onto
+			// view/like/share records — undeclared columns on those event
+			// types (schema rule 1). Viral users virtually always carry all
+			// three templates, so skipping is behavior-neutral in practice.
+			const viewTemplate = userEvents.find(e => e.event === "post viewed");
+			const likeTemplate = userEvents.find(e => e.event === "post liked");
+			const shareTemplate = userEvents.find(e => e.event === "post shared");
+
+			if (viewTemplate) for (let i = 0; i < viralViews; i++) {
+				injected.push({
+					...viewTemplate,
+					event: "post viewed",
+					time: eventTime.add(chance.integer({ min: 1, max: 180 }), 'minutes').toISOString(),
+					user_id: event.user_id,
+					post_type: event.post_type || "text",
+					source: chance.pickone(["feed", "explore", "search"]),
+					view_duration_sec: chance.integer({ min: 5, max: 90 }),
+				});
+			}
+			if (likeTemplate) for (let i = 0; i < viralLikes; i++) {
+				injected.push({
+					...likeTemplate,
+					event: "post liked",
+					time: eventTime.add(chance.integer({ min: 2, max: 240 }), 'minutes').toISOString(),
+					user_id: event.user_id,
+					post_type: event.post_type || "text",
+				});
+			}
+			if (shareTemplate) for (let i = 0; i < viralShares; i++) {
+				injected.push({
+					...shareTemplate,
+					event: "post shared",
+					time: eventTime.add(chance.integer({ min: 5, max: 300 }), 'minutes').toISOString(),
+					user_id: event.user_id,
+					share_destination: chance.pickone(["repost", "dm", "external", "copy_link"]),
+				});
+			}
+
+			userEvents.splice(idx + 1, 0, ...injected);
+		}
+
+		// H2: FOLLOW-BACK SNOWBALL — extra post + comment per user-followed cluster.
+		// Discovery: cohort users with >=5 user-followed events, compare posts/user.
+		if (followReceivedCount >= FOLLOW_SNOWBALL_THRESHOLD && event.event === "post created") {
+			if (chance.bool({ likelihood: FOLLOW_SNOWBALL_LIKELIHOOD })) {
+				const commentTemplate = userEvents.find(e => e.event === "comment posted");
+				const duplicatePost = {
+					...event,
+					time: eventTime.add(chance.integer({ min: 30, max: 240 }), 'minutes').toISOString(),
+					user_id: event.user_id,
+					post_type: chance.pickone(["text", "image", "video"]),
+					character_count: chance.integer({ min: 10, max: 280 }),
+					has_media: chance.bool({ likelihood: 60 }),
+					hashtag_count: chance.integer({ min: 0, max: 5 }),
+				};
+				const extraComment = {
+					...(commentTemplate || event),
+					event: "comment posted",
+					time: eventTime.add(chance.integer({ min: 10, max: 120 }), 'minutes').toISOString(),
+					user_id: event.user_id,
+					comment_length: chance.integer({ min: 5, max: 200 }),
+					has_mention: chance.bool({ likelihood: 40 }),
+				};
+				userEvents.splice(idx + 1, 0, duplicatePost, extraComment);
+			}
+		}
+
+		// H6: CREATOR MONETIZATION — 3x post/story rate for subscribers.
+		// Discovery: cohort users with creator-subscription-started event, compare posts/user.
+		if (hasCreatorSubscription && event.event === "post created") {
+			for (let i = 0; i < CREATOR_POST_CLONES; i++) {
+				const extraPost = {
+					...event,
+					time: eventTime.add(chance.integer({ min: 1, max: 12 }), 'hours').toISOString(),
+					user_id: event.user_id,
+					post_type: chance.pickone(["text", "image", "video", "link"]),
+					character_count: chance.integer({ min: 20, max: 280 }),
+					has_media: chance.bool({ likelihood: 70 }),
+					hashtag_count: chance.integer({ min: 1, max: 8 }),
+				};
+				userEvents.splice(idx + 1, 0, extraPost);
+			}
+		}
+		if (hasCreatorSubscription && event.event === "story created") {
+			for (let i = 0; i < CREATOR_STORY_CLONES; i++) {
+				const extraStory = {
+					...event,
+					time: eventTime.add(chance.integer({ min: 1, max: 8 }), 'hours').toISOString(),
+					user_id: event.user_id,
+					story_type: chance.pickone(["photo", "video", "text"]),
+					has_filter: chance.bool({ likelihood: 60 }),
+					has_sticker: chance.bool({ likelihood: 40 }),
+				};
+				userEvents.splice(idx + 1, 0, extraStory);
+			}
+		}
+		if (hasCreatorSubscription && event.event === "post viewed") {
+			if (chance.bool({ likelihood: CREATOR_VIEW_CLONE_LIKELIHOOD })) {
+				const analyticsView = {
+					...event,
+					time: eventTime.add(chance.integer({ min: 1, max: 30 }), 'minutes').toISOString(),
+					user_id: event.user_id,
+					post_type: event.post_type || "text",
+					source: "profile",
+					view_duration_sec: chance.integer({ min: 10, max: 60 }),
+				};
+				userEvents.splice(idx + 1, 0, analyticsView);
+			}
+		}
+	}
+
+	// H8: WEEKEND CONTENT SURGE — duplicate weekend posts/stories with offset.
+	// Discovery: line chart by day-of-week shows Sat/Sun bump.
+	for (let idx = userEvents.length - 1; idx >= 0; idx--) {
+		const event = userEvents[idx];
+		if (event.event === "post created" || event.event === "story created") {
+			const dow = new Date(event.time).getUTCDay();
+			if ((dow === 0 || dow === 6) && chance.bool({ likelihood: WEEKEND_CLONE_LIKELIHOOD })) {
+				const etime = dayjs(event.time);
+				const dup = {
+					...event,
+					time: etime.add(chance.integer({ min: 1, max: 3 }), 'hours').toISOString(),
+				};
+				userEvents.splice(idx + 1, 0, dup);
+			}
+		}
+	}
+
+	// H3: ALGORITHM CHANGE — day 45 flips feed → explore on post viewed.
+	// H4: ENGAGEMENT BAIT — 20% of post-viewed events get crushed duration.
+	// H5: NOTIFICATION RE-ENGAGEMENT — after day 30, 30% of views → notification.
+	// All three run AFTER injection passes so they apply to cloned events too.
+	const algorithmChangeDay = datasetStart.add(ALGORITHM_CHANGE_DAY, 'days');
+	const reengagementStart = datasetStart.add(REENGAGEMENT_START_DAY, 'days');
+	userEvents.forEach(e => {
+		if (e.event === "post viewed") {
+			const eventTime = dayjs(e.time);
+
+			// H3: Algorithm Change
+			if (eventTime.isAfter(algorithmChangeDay)) {
+				if (chance.bool({ likelihood: ALGORITHM_FLIP_LIKELIHOOD })) {
+					e.source = "explore";
+				}
+			} else {
+				if (chance.bool({ likelihood: ALGORITHM_FLIP_LIKELIHOOD })) {
+					e.source = "feed";
+				}
+			}
+
+			// H4: Engagement Bait — 20% crushed view duration
+			if (chance.bool({ likelihood: ENGAGEMENT_BAIT_LIKELIHOOD })) {
+				e.view_duration_sec = chance.integer({ min: 1, max: 5 });
+			}
+
+			// H5: Notification Re-engagement (runs after #3 so can override)
+			if (eventTime.isAfter(reengagementStart) && chance.bool({ likelihood: REENGAGEMENT_LIKELIHOOD })) {
+				e.source = "notification";
+			}
+		}
+	});
+
+	// H7: TOXICITY CHURN — drop 60% of activity after day 30 for high reporters.
+	// Discovery: cohort users with >=2 report-submitted events, observe retention drop.
+	if (reportSubmittedCount >= TOXICITY_THRESHOLD) {
+		const churnCutoff = datasetStart.add(TOXICITY_CUTOFF_DAYS, 'days');
+		for (let i = userEvents.length - 1; i >= 0; i--) {
+			const evt = userEvents[i];
+			if (dayjs(evt.time).isAfter(churnCutoff) && chance.bool({ likelihood: TOXICITY_DROP_LIKELIHOOD })) {
+				userEvents.splice(i, 1);
+			}
+		}
+	}
+
+	// H9: POST-CREATED MAGIC NUMBER (no flags)
+	// Sweet 3-7 posts → +40% on comment_length on the user's comment events.
+	// Over 8+ → -30% comment_length (engagement burnout).
+	if (postCreatedCount >= POST_SWEET_MIN && postCreatedCount <= POST_SWEET_MAX) {
+		userEvents.forEach(e => {
+			if (e.event === 'comment posted' && typeof e.comment_length === 'number') {
+				e.comment_length = Math.round(e.comment_length * POST_SWEET_COMMENT_BOOST);
+			}
+		});
+	} else if (postCreatedCount >= POST_OVER_THRESHOLD) {
+		userEvents.forEach(e => {
+			if (e.event === 'comment posted' && typeof e.comment_length === 'number') {
+				e.comment_length = Math.round(e.comment_length * POST_OVER_COMMENT_FACTOR);
+			}
+		});
+	}
+
+	return record;
+}
+
+// ── CONFIG ──
+/** @type {Config} */
+const config = {
+	version: 2,
+	seed: SEED,
+	datasetStart: DATASET_START,
+	datasetEnd: DATASET_END,
+	// All weights >= 1.0 are accept-always in TimeSoup's rejection sampler →
+	// flat organic DOW. Load-bearing: removing this key re-activates
+	// DEFAULT_DOW_WEIGHTS (weekend-suppressed), which would fight H8.
+	soup: { dayOfWeekWeights: [1.0, 1.0, 1.0, 1.0, 1.0, 1.2, 1.2] },
+	avgEventsPerUserPerDay: EVENTS_PER_DAY,
+	numUsers: NUM_USERS,
+	format: "json",
+	gzip: true,
+	credentials: {
+		token,
+	},
+	switches: {
+		hasSessionIds: true,
+		alsoInferFunnels: false,
+		hasLocation: true,
+		hasAndroidDevices: true,
+		hasIOSDevices: true,
+		hasDesktopDevices: true,
+		hasBrowser: false,
+		hasCampaigns: false,
+		isAnonymous: false,
+		hasAdSpend: false,
+		hasAvatar: true,
+	},
+	identity: {
+		avgDevicePerUser: 2,
+	},
+	concurrency: 1,
+	writeToDisk: false,
+	scdProps: {
+		account_type: {
+			values: ["personal", "creator", "business", "verified"],
+			frequency: "month",
+			timing: "fuzzy",
+			max: 6
+		},
+		community_status: {
+			values: ["new", "growing", "established", "featured"],
+			frequency: "month",
+			timing: "fixed",
+			max: 6,
+			type: "community_id"
+		}
+	},
+
+	funnels: [
+		{
+			sequence: ["account created", "profile updated", "post created"],
+			isFirstFunnel: true,
+			conversionRate: 70,
+			timeToConvert: 0.5,
+		},
+		{
+			// Feed consumption: view → like → comment (most common loop)
+			sequence: ["post viewed", "post liked", "comment posted"],
+			conversionRate: 45,
+			timeToConvert: 0.5,
+			weight: 6,
+		},
+		{
+			// Content creation cycle: create → views → engagement
+			sequence: ["post created", "post viewed", "post liked", "post shared"],
+			conversionRate: 30,
+			timeToConvert: 3,
+			weight: 3,
+		},
+		{
+			// Stories engagement
+			sequence: ["story created", "story viewed", "dm sent"],
+			conversionRate: 40,
+			timeToConvert: 1,
+			weight: 3,
+		},
+		{
+			// Discovery and follow loop
+			sequence: ["search performed", "post viewed", "user followed"],
+			conversionRate: 35,
+			timeToConvert: 1,
+			weight: 2,
+		},
+		{
+			// Notifications driving re-engagement
+			sequence: ["notification received", "post viewed", "post liked"],
+			conversionRate: 50,
+			timeToConvert: 0.5,
+			weight: 2,
+		},
+		{
+			// Profile management and creator monetization
+			sequence: ["profile updated", "creator subscription started", "post created"],
+			conversionRate: 15,
+			timeToConvert: 24,
+			weight: 1,
+		},
+		{
+			// Ad interaction and moderation
+			sequence: ["ad viewed", "ad clicked", "report submitted"],
+			conversionRate: 20,
+			timeToConvert: 2,
+			weight: 1,
+		},
+	],
+
+	events: [
+		{
+			event: "account created",
+			weight: 1,
+			isFirstEvent: true,
+			isAuthEvent: true,
+			properties: {
+				"signup_method": ["email", "google", "apple", "sso"],
+				"referred_by": ["organic", "friend", "ad", "influencer"],
+			}
+		},
+		{
+			event: "post created",
+			weight: 12,
+			isStrictEvent: false,
+			properties: {
+				"post_type": ["text", "image", "video", "poll", "link"],
+				"character_count": u.weighNumRange(1, 280),
+				"has_media": [false, false, false, true, true],
+				"hashtag_count": u.weighNumRange(0, 10, 0.5),
+			}
+		},
+		{
+			event: "post viewed",
+			weight: 30,
+			isStrictEvent: false,
+			properties: {
+				"post_type": ["text", "image", "video", "poll", "link"],
+				"view_duration_sec": u.weighNumRange(1, 120, 0.3, 5),
+				"source": ["feed", "explore", "search", "profile", "notification"],
+			}
+		},
+		{
+			event: "post liked",
+			weight: 18,
+			isStrictEvent: false,
+			properties: {
+				"post_type": ["text", "image", "video", "poll", "link"],
+			}
+		},
+		{
+			event: "post shared",
+			weight: 6,
+			isStrictEvent: false,
+			properties: {
+				"share_destination": ["repost", "dm", "external", "copy_link"],
+			}
+		},
+		{
+			event: "comment posted",
+			weight: 10,
+			isStrictEvent: false,
+			properties: {
+				"comment_length": u.weighNumRange(1, 500, 0.3, 20),
+				"has_mention": [true, false, false],
+			}
+		},
+		{
+			event: "user followed",
+			weight: 8,
+			isStrictEvent: false,
+			properties: {
+				"discovery_source": ["suggested", "search", "post", "profile", "mutual"],
+			}
+		},
+		{
+			event: "user unfollowed",
+			weight: 2,
+			properties: {
+				"reason": ["content_quality", "too_frequent", "lost_interest", "offensive"],
+			}
+		},
+		{
+			event: "story viewed",
+			weight: 15,
+			isStrictEvent: false,
+			properties: {
+				"story_type": ["photo", "video", "text"],
+				"view_duration_sec": u.weighNumRange(1, 30, 0.5, 5),
+				"completed": [false, false, true, true, true],
+			}
+		},
+		{
+			event: "story created",
+			weight: 5,
+			isStrictEvent: false,
+			properties: {
+				"story_type": ["photo", "video", "text"],
+				"has_filter": [true, false],
+				"has_sticker": [false, false, true],
+			}
+		},
+		{
+			event: "search performed",
+			weight: 7,
+			properties: {
+				"search_type": ["users", "hashtags", "posts"],
+				"results_count": u.weighNumRange(0, 50, 0.5, 10),
+			}
+		},
+		{
+			event: "notification received",
+			weight: 12,
+			properties: {
+				"notification_type": ["like", "follow", "comment", "mention", "trending"],
+				"clicked": [false, false, false, true, true],
+			}
+		},
+		{
+			event: "dm sent",
+			weight: 8,
+			isStrictEvent: false,
+			properties: {
+				"message_type": ["text", "image", "voice", "link"],
+				"conversation_length": u.weighNumRange(1, 100),
+			}
+		},
+		{
+			event: "ad viewed",
+			weight: 10,
+			properties: {
+				"ad_format": ["feed_native", "story", "banner", "video"],
+				"ad_category": ["retail", "tech", "food", "finance", "entertainment"],
+				"view_duration_sec": u.weighNumRange(1, 30, 0.3),
+			}
+		},
+		{
+			event: "ad clicked",
+			weight: 2,
+			properties: {
+				"ad_format": ["feed_native", "story", "banner", "video"],
+				"ad_category": ["retail", "tech", "food", "finance", "entertainment"],
+			}
+		},
+		{
+			event: "report submitted",
+			weight: 1,
+			properties: {
+				"report_type": ["spam", "harassment", "misinformation", "hate_speech", "other"],
+				"content_type": ["post", "comment", "user", "dm"],
+			}
+		},
+		{
+			event: "profile updated",
+			weight: 3,
+			properties: {
+				"field_updated": ["bio", "avatar", "display_name", "privacy_settings", "interests"],
+			}
+		},
+		{
+			event: "creator subscription started",
+			weight: 2,
+			properties: {
+				"tier": ["basic", "premium", "vip"],
+				"price_usd": [4.99, 9.99, 19.99],
+			}
+		},
+	],
+
+	superProps: {
+		app_version: ["4.0", "4.1", "4.2", "4.3", "5.0"],
+		account_type: ["personal", "creator", "business"],
+	},
+
+	userProps: {
+		app_version: ["4.0", "4.1", "4.2", "4.3", "5.0"],
+		account_type: ["personal", "creator", "business"],
+		"follower_count": u.weighNumRange(0, 10000, 0.2, 50),
+		"following_count": u.weighNumRange(0, 5000, 0.3, 100),
+		"bio_length": u.weighNumRange(0, 160),
+		"verified": [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true],
+		"content_niche": ["lifestyle", "tech", "food", "fitness", "travel", "comedy", "news", "art"],
+	},
+
+	groupKeys: [
+		["community_id", 100, ["post created", "comment posted", "post liked", "post shared"]],
+	],
+
+	groupProps: {
+		community_id: {
+			"name": () => `${chance.word()} ${chance.pickone(["Hub", "Circle", "Squad", "Zone", "Space"])}`,
+			"member_count": u.weighNumRange(50, 5000, 0.3, 200),
+			"category": ["technology", "entertainment", "sports", "politics", "art", "science"],
+			"is_moderated": [false, false, false, true, true, true, true, true, true, true],
+		}
+	},
+
+	lookupTables: [],
+
+	hook(record, type, meta) {
+		if (type === "funnel-post") return handleFunnelPostHooks(record, meta);
+		if (type === "everything") return handleEverythingHooks(record, meta);
+		return record;
+	}
+};
+
+export default config;
+
+// ── STORIES (verification contract — consumed by social.verify.mjs) ──
+/*
+ * DERIVATION NOTES (all numbers measured at 2K reduced scale, seed
+ * harness-social, plus an exact organic counterfactual run — same seed,
+ * identity hook — used to separate engineered effects from the organic
+ * activity curve):
+ *
+ *   - H1: top-3% of users hold 0.623 of all post viewed/liked/shared
+ *     events (organic counterfactual: 0.057). 59 users show 500+
+ *     engagement events vs a 45 median among 10+ posters. The old
+ *     "10+ post created cohort" read is dead: 85% of users clear 10
+ *     final posts (organic median ~10, clones inflate further).
+ *   - H2: big-follower (5+ user followed) posts/user 42.30 vs 28.19 =
+ *     1.50x. Decomposition via counterfactual: 1.24x activity confound
+ *     x 1.21x true hook increment (H6's +2-clones-per-post dilutes
+ *     H2's +50%-of-organic-posts for subscriber users).
+ *   - H3+H5 regimes (post viewed source shares): pre-d30 feed 0.779,
+ *     notification 0.033; day 30-45 feed 0.545, notification 0.320;
+ *     post-d45 explore 0.539, notification 0.322, feed 0.055.
+ *   - H4: share of views <= 5 sec = 0.204 (organic 0.000 — generator
+ *     floors organic durations above 5s; the low mode is 100% hook-made).
+ *   - H6: subscribers = 78% of users (weight-2 event over 121 days).
+ *     posts/user sub vs non 35.34/10.63 = 3.32; stories/user 32.32/8.47
+ *     = 3.82 (purer read: H2 dupes lift non-sub posts but not stories).
+ *     Paired lifts: sub posts 3.07x, stories 3.00x; non posts 1.24x.
+ *   - H7: post/pre-day-30 event ratio reporters (2+ final reports) 1.27
+ *     vs normal 3.21 → 0.395 contrast. Paired per-user drop 0.419
+ *     (60% knob). Growth shape makes the raw post/pre ~3.2 for everyone.
+ *   - H8: weekend/weekday daily creations hooked 1.2545, organic 0.9802
+ *     (soup weights >= 1.0 are accept-always → flat DOW; the old "soup
+ *     dampens weekends ~0.55x" claim in the v1.5 verifier was wrong).
+ *   - H9 (final-count buckets — analyst view): sweet(3-7) avg
+ *     comment_length 250.09 (162 users), low(0-2) 200.28 (33 users),
+ *     over(8+) 145.43 (1768 users) → sweet/over 1.72, ordering
+ *     sweet > low > over. Paired per-user lifts land on the knobs:
+ *     low 0.998, sweet 1.397, over 0.681.
+ *   - H10 (emulateBreakdown timeToConvert @6h, stable 1h-24h): personal
+ *     median 22.9min (n=64), creator 16.3 (51), business 16.2 (47) →
+ *     fast/personal 0.709. Organic counterfactual: 0.963 baseline ratio,
+ *     medians ~20min. Per-leg attenuation: 0.71 knob → 0.81 realized,
+ *     1.25 → 1.11 (greedy step-matching, unique first-step anchor).
+ *   - Identity: uid share 1.0 (auth on first event), device share
+ *     0.9989, devices/user 2.07 (avgDevicePerUser: 2).
+ *
+ * Scale guards sit at ~50% of expected 10K populations, so 2K runs trip
+ * WEAK by design; verdicts ship only from full-fidelity runs.
+ *
+ * Fix-round Q5 (2026-07-04, adversarial-review S1): NAILED bands
+ * re-derived as knob ±10% wherever the knob converts directly to the
+ * metric (H6 3.0x clone lifts, H9 sweet/over 1.4/0.7 → 2.00 + margin
+ * knobs, H10 TTC 0.71/1.25 → 0.568; H4/H7/dpu tightened to ±10%).
+ * Where the realized magnitude is confounded (H1 concentration share,
+ * H5 pre-d30 clone dilution), the assertion is a knob-bounded
+ * floor/ceiling that grades STRONG by design. The measured numbers
+ * above remain as documentation of expected full-fidelity readings —
+ * they are no longer verdict anchors.
+ */
+
+const EV = `read_json_auto('{{PREFIX}}-EVENTS*.json', sample_size=-1, union_by_name=true)`;
+
+const bandVerdict = (x, nailed, strong, detail, inverse = () => false) => {
+	if (x == null || Number.isNaN(Number(x))) return { verdict: "NONE", detail: `${detail} — metric missing` };
+	const v = Number(x);
+	if (inverse(v)) return { verdict: "INVERSE", detail };
+	if (v >= nailed[0] && v <= nailed[1]) return { verdict: "NAILED", detail };
+	if (v >= strong[0] && v <= strong[1]) return { verdict: "STRONG", detail };
+	return { verdict: "WEAK", detail };
+};
+const guarded = (ok, detail, inner) => ok ? inner() : { verdict: "WEAK", detail: `${detail} — cohort below scale guard (expected at reduced scale)` };
+const worstOf = (...verdicts) => { const order = ["INVERSE", "NONE", "WEAK", "STRONG", "NAILED"]; return order.find(o => verdicts.some(v => v.verdict === o)) || "NONE"; };
+const cellsOf = (rows, key) => Object.fromEntries((rows || []).map(r => [r[key], r]));
+
+const DAY30 = `TIMESTAMP '2026-01-31 00:00:00'`;
+const DAY45 = `TIMESTAMP '2026-02-15 00:00:00'`;
+
+const SOURCE_WINDOWS_SQL = `SELECT
+  CASE WHEN time::TIMESTAMP < ${DAY30} THEN 'pre_d30'
+       WHEN time::TIMESTAMP < ${DAY45} THEN 'd30_45'
+       ELSE 'post_d45' END AS win,
+  COUNT(*) AS n,
+  AVG((source = 'feed')::INT) AS feed,
+  AVG((source = 'explore')::INT) AS explore,
+  AVG((source = 'notification')::INT) AS notif
+FROM ${EV} WHERE event = 'post viewed' GROUP BY 1 ORDER BY 1`;
+
+export const stories = [
+	{
+		id: "social-h1-viral-concentration",
+		hook: "H1",
+		archetype: "cohort-count-scale",
+		narrative: "Viral creators (~3% of users) dominate engagement: top-3% share floor 0.35 (~6x the 0.057 organic ceiling; magnitude confounded — graded STRONG by design).",
+		assertions: [
+			{
+				breakdown: {
+					type: "duckdb",
+					sql: `WITH pu AS (
+  SELECT user_id::VARCHAR AS uid,
+    COUNT(*) FILTER (WHERE event IN ('post viewed', 'post liked', 'post shared')) AS eng
+  FROM ${EV} WHERE user_id IS NOT NULL GROUP BY 1
+), ranked AS (
+  SELECT eng, ROW_NUMBER() OVER (ORDER BY eng DESC) AS rn,
+    COUNT(*) OVER () AS n, SUM(eng) OVER () AS tot
+  FROM pu
+)
+SELECT SUM(eng)::DOUBLE / MAX(tot) AS top3_share, MAX(n) AS users
+FROM ranked WHERE rn <= CEIL(n * 0.03)`,
+				},
+				assert: (rows) => {
+					const r = rows?.[0] || {};
+					return guarded(Number(r.users) >= 5000, `users=${r.users ?? 0}`, () => {
+						const share = Number(r.top3_share);
+						const detail = `top-3% share of engagement events=${share.toFixed(3)} (measured 0.623; organic counterfactual 0.057)`;
+						// Fix-round Q5 (S1): concentration magnitude is confounded (organic engagement
+						// volume, H6 view clones, H7 drops) — not derivable from VIRAL_* knobs. Knob-derived
+						// floor: each viral creator injects >= 10 posts × 3 types × VIRAL_CLONES_MIN(60)
+						// = 1,800 engagement events vs ~145 organic events/user, so share >= 0.35 (~6× the
+						// 0.057 organic ceiling) grades STRONG by design; no NAILED band from a measured value.
+						if (share <= 0.15) return { verdict: "INVERSE", detail };
+						if (share >= 0.35) return { verdict: "STRONG", detail };
+						return { verdict: "WEAK", detail };
+					});
+				},
+			},
+		],
+	},
+	{
+		id: "social-h2-followback-snowball",
+		hook: "H2",
+		archetype: "cohort-count-scale",
+		narrative: "Users with 5+ user-followed events post ~1.5x more (1.24x activity confound x 1.21x hook increment).",
+		assertions: [
+			{
+				breakdown: {
+					type: "duckdb",
+					sql: `WITH pu AS (
+  SELECT user_id::VARCHAR AS uid,
+    COUNT(*) FILTER (WHERE event = 'user followed') AS fc,
+    COUNT(*) FILTER (WHERE event = 'post created') AS pc
+  FROM ${EV} WHERE user_id IS NOT NULL GROUP BY 1
+)
+SELECT (fc >= 5) AS big, COUNT(*) AS users, AVG(pc) AS posts
+FROM pu GROUP BY 1 ORDER BY 1`,
+				},
+				assert: (rows) => {
+					const by = cellsOf(rows, "big");
+					const big = by.true, small = by.false;
+					return guarded(Number(big?.users) >= 570 && Number(small?.users) >= 4000, `cohorts: big=${big?.users ?? 0} small=${small?.users ?? 0}`, () => {
+						const ratio = Number(big.posts) / Number(small.posts);
+						const detail = `posts/user 5+ follows vs rest=${ratio.toFixed(3)} (measured 1.501; ${Number(big.posts).toFixed(1)} vs ${Number(small.posts).toFixed(1)})`;
+						return bandVerdict(ratio, [1.40, 1.62], [1.28, 1.80], detail, v => v <= 1.0);
+					});
+				},
+			},
+		],
+	},
+	{
+		id: "social-h3-algorithm-change",
+		hook: "H3",
+		archetype: "temporal-inflection",
+		narrative: "Day-45 feed→explore flip: pre-d30 feed 78% of views; post-d45 explore 54%, feed collapses to 5%.",
+		assertions: [
+			{
+				breakdown: { type: "duckdb", sql: SOURCE_WINDOWS_SQL },
+				assert: (rows) => {
+					const by = cellsOf(rows, "win");
+					const pre = by.pre_d30, post = by.post_d45;
+					return guarded(Number(pre?.n) >= 90000 && Number(post?.n) >= 230000, `views: pre_d30=${pre?.n ?? 0} post_d45=${post?.n ?? 0}`, () => {
+						const preFeed = Number(pre.feed), postExplore = Number(post.explore), postFeed = Number(post.feed);
+						const detail = `pre-d30 feed=${preFeed.toFixed(3)} (measured 0.779); post-d45 explore=${postExplore.toFixed(3)} (0.539), feed=${postFeed.toFixed(3)} (0.055)`;
+						const legs = [
+							bandVerdict(preFeed, [0.74, 0.82], [0.70, 0.86], detail, v => v <= 0.30),
+							bandVerdict(postExplore, [0.50, 0.58], [0.45, 0.63], detail, v => v <= 0.15),
+							// postFeed width is mixture-motivated (H1 clones carry 1/3 feed pre-flip); the
+							// pure knob residual 0.30 × 0.70 × 0.20 = 0.042 sits inside the NAILED band.
+							bandVerdict(postFeed, [0.03, 0.08], [0.02, 0.12], detail, v => v >= 0.30),
+						];
+						return { verdict: worstOf(...legs), detail };
+					});
+				},
+			},
+		],
+	},
+	{
+		id: "social-h4-engagement-bait",
+		hook: "H4",
+		archetype: "composition-drift",
+		narrative: "20% of post-viewed events get view_duration_sec crushed to 1-5s — a bimodal low mode that is 100% hook-made.",
+		assertions: [
+			{
+				breakdown: {
+					type: "duckdb",
+					sql: `SELECT COUNT(*) AS n, AVG((view_duration_sec <= 5)::INT) AS crushed_share
+FROM ${EV} WHERE event = 'post viewed' AND view_duration_sec IS NOT NULL`,
+				},
+				assert: (rows) => {
+					const r = rows?.[0] || {};
+					return guarded(Number(r.n) >= 375000, `views=${r.n ?? 0}`, () => {
+						const share = Number(r.crushed_share);
+						const detail = `share of views <= 5 sec=${share.toFixed(4)} (measured 0.204; organic 0.000, generator floors durations above 5s)`;
+						// Fix-round Q5 (S1): NAILED = ENGAGEMENT_BAIT_LIKELIHOOD(20) → 0.20 ±10%.
+						return bandVerdict(share, [0.18, 0.22], [0.15, 0.26], detail, v => v <= 0.05);
+					});
+				},
+			},
+		],
+	},
+	{
+		id: "social-h5-notification-reengagement",
+		hook: "H5",
+		archetype: "temporal-inflection",
+		narrative: "After day 30, ~32% of post-viewed events carry source=notification (knob 0.30 + residual); pre-d30 share stays under the 0.066 knob ceiling (clone-diluted — pre leg graded STRONG by design).",
+		assertions: [
+			{
+				breakdown: { type: "duckdb", sql: SOURCE_WINDOWS_SQL },
+				assert: (rows) => {
+					const by = cellsOf(rows, "win");
+					const pre = by.pre_d30, mid = by.d30_45, post = by.post_d45;
+					const postN = Number(mid?.n ?? 0) + Number(post?.n ?? 0);
+					return guarded(Number(pre?.n) >= 90000 && postN >= 280000, `views: pre_d30=${pre?.n ?? 0} post_d30=${postN}`, () => {
+						const preNotif = Number(pre.notif);
+						const postNotif = (Number(mid.notif) * Number(mid.n) + Number(post.notif) * Number(post.n)) / postN;
+						const detail = `notification share pre-d30=${preNotif.toFixed(4)} (measured 0.033) → post-d30=${postNotif.toFixed(4)} (measured 0.322)`;
+						// Fix-round Q5 (S1): pre-d30 knob-implied notif share = 0.30 unflipped-by-H3
+						// × 0.20 organic = 0.06; H1/H6 clone dilution (clones never stamp notification)
+						// drags it lower by a non-knob-derivable amount. Knob-derived ceiling (low-before
+						// check): <= 0.066 (knob + 10%) grades STRONG by design; no NAILED band. postNotif
+						// keeps its knob band (REENGAGEMENT_LIKELIHOOD=30 floor + 0.042 residual).
+						const legPre = preNotif >= 0.15 ? { verdict: "INVERSE", detail }
+							: preNotif <= 0.066 ? { verdict: "STRONG", detail }
+							: { verdict: "WEAK", detail };
+						const legs = [
+							legPre,
+							bandVerdict(postNotif, [0.29, 0.35], [0.26, 0.38], detail, v => v <= 0.08),
+						];
+						return { verdict: worstOf(...legs), detail };
+					});
+				},
+			},
+		],
+	},
+	{
+		id: "social-h6-creator-monetization",
+		hook: "H6",
+		archetype: "cohort-count-scale",
+		narrative: "Subscribers (~78% of users) post and story ~3.0x more per user (CREATOR_*_CLONES=2 knob; activity confound lifts measured readings to 3.3x/3.8x — graded STRONG by design).",
+		assertions: [
+			{
+				breakdown: {
+					type: "duckdb",
+					sql: `WITH pu AS (
+  SELECT user_id::VARCHAR AS uid,
+    BOOL_OR(event = 'creator subscription started') AS sub,
+    COUNT(*) FILTER (WHERE event = 'post created') AS pc,
+    COUNT(*) FILTER (WHERE event = 'story created') AS sc
+  FROM ${EV} WHERE user_id IS NOT NULL GROUP BY 1
+)
+SELECT sub, COUNT(*) AS users, AVG(pc) AS posts, AVG(sc) AS stories
+FROM pu GROUP BY 1 ORDER BY 1`,
+				},
+				assert: (rows) => {
+					const by = cellsOf(rows, "sub");
+					const s = by.true, n = by.false;
+					return guarded(Number(s?.users) >= 3800 && Number(n?.users) >= 1100, `cohorts: sub=${s?.users ?? 0} non=${n?.users ?? 0}`, () => {
+						const postRatio = Number(s.posts) / Number(n.posts);
+						const storyRatio = Number(s.stories) / Number(n.stories);
+						const detail = `sub vs non posts/user=${postRatio.toFixed(3)} (measured 3.32); stories/user=${storyRatio.toFixed(3)} (measured 3.82)`;
+						// Fix-round Q5 (S1): knob is CREATOR_POST_CLONES=2 / CREATOR_STORY_CLONES=2 → 3.0x
+						// per-user lift for both metrics; NAILED = knob ±10% = [2.70, 3.30]. Measured 3.32
+						// (posts) and 3.82 (stories) carry a subscriber activity confound on top of the knob
+						// (paired lifts land on 3.07x/3.00x) — those readings grade STRONG by design.
+						const legs = [
+							bandVerdict(postRatio, [2.70, 3.30], [2.40, 4.00], detail, v => v <= 1.5),
+							bandVerdict(storyRatio, [2.70, 3.30], [2.40, 4.60], detail, v => v <= 1.5),
+						];
+						return { verdict: worstOf(...legs), detail };
+					});
+				},
+			},
+		],
+	},
+	{
+		id: "social-h7-toxicity-churn",
+		hook: "H7",
+		archetype: "retention-divergence",
+		narrative: "Users with 2+ reports lose 60% of post-day-30 activity: post/pre ratio 1.27 vs 3.21 for the rest (0.40x contrast).",
+		assertions: [
+			{
+				breakdown: {
+					type: "duckdb",
+					sql: `WITH pu AS (
+  SELECT user_id::VARCHAR AS uid,
+    COUNT(*) FILTER (WHERE event = 'report submitted') AS rep,
+    COUNT(*) FILTER (WHERE time::TIMESTAMP <= ${DAY30}) AS pre,
+    COUNT(*) FILTER (WHERE time::TIMESTAMP > ${DAY30}) AS post
+  FROM ${EV} WHERE user_id IS NOT NULL GROUP BY 1
+)
+SELECT (rep >= 2) AS reporter, COUNT(*) AS users, AVG(post::DOUBLE / pre) AS ratio
+FROM pu WHERE pre > 0 GROUP BY 1 ORDER BY 1`,
+				},
+				assert: (rows) => {
+					const by = cellsOf(rows, "reporter");
+					const rep = by.true, norm = by.false;
+					return guarded(Number(rep?.users) >= 310 && Number(norm?.users) >= 4000, `cohorts: reporters=${rep?.users ?? 0} normal=${norm?.users ?? 0}`, () => {
+						const contrast = Number(rep.ratio) / Number(norm.ratio);
+						const detail = `post/pre-d30 ratio reporters=${Number(rep.ratio).toFixed(2)} vs normal=${Number(norm.ratio).toFixed(2)} → contrast=${contrast.toFixed(3)} (measured 0.395; 60% drop knob)`;
+						// Fix-round Q5 (S1): NAILED = survival knob (1 − TOXICITY_DROP_LIKELIHOOD(60)/100)
+						// = 0.40 ±10% (was ±15%).
+						return bandVerdict(contrast, [0.36, 0.44], [0.28, 0.55], detail, v => v >= 0.85);
+					});
+				},
+			},
+		],
+	},
+	{
+		id: "social-h8-weekend-surge",
+		hook: "H8",
+		archetype: "bespoke",
+		narrative: "Sat/Sun creations run ~1.25x weekday daily rate on a flat organic DOW baseline (0.98 counterfactual).",
+		assertions: [
+			{
+				breakdown: {
+					type: "duckdb",
+					sql: `WITH d AS (
+  SELECT time::DATE AS dt, COUNT(*) AS n FROM ${EV}
+  WHERE event IN ('post created', 'story created') GROUP BY 1
+)
+SELECT
+  AVG(n) FILTER (WHERE EXTRACT(DOW FROM dt) IN (0, 6)) AS wkn,
+  AVG(n) FILTER (WHERE EXTRACT(DOW FROM dt) NOT IN (0, 6)) AS wkd
+FROM d`,
+				},
+				assert: (rows) => {
+					const r = rows?.[0] || {};
+					return guarded(Number(r.wkd) >= 2300, `weekday creations/day=${r.wkd ? Number(r.wkd).toFixed(0) : 0}`, () => {
+						const ratio = Number(r.wkn) / Number(r.wkd);
+						const detail = `weekend/weekday daily creations=${ratio.toFixed(3)} (measured 1.254; organic counterfactual 0.980 — flat DOW)`;
+						return bandVerdict(ratio, [1.19, 1.32], [1.10, 1.42], detail, v => v <= 1.02);
+					});
+				},
+			},
+		],
+	},
+	{
+		id: "social-h9-magic-number",
+		hook: "H9",
+		archetype: "frequency-sweet-spot",
+		narrative: "Sweet-spot posters (3-7 final posts) out-comment 8+ posters ~2.0x per knobs (1.4 boost / 0.7 factor); bucket migration attenuates the measured ratio to ~1.7 — graded STRONG by design. Ordering sweet > low > over.",
+		assertions: [
+			{
+				breakdown: {
+					type: "duckdb",
+					sql: `WITH pu AS (
+  SELECT user_id::VARCHAR AS uid, COUNT(*) FILTER (WHERE event = 'post created') AS pc
+  FROM ${EV} WHERE user_id IS NOT NULL GROUP BY 1
+)
+SELECT CASE WHEN pu.pc >= 8 THEN 'over' WHEN pu.pc >= 3 THEN 'sweet' ELSE 'low' END AS bucket,
+  COUNT(DISTINCT pu.uid) AS users, COUNT(*) AS n_comments, AVG(e.comment_length) AS len
+FROM ${EV} e JOIN pu ON e.user_id::VARCHAR = pu.uid
+WHERE e.event = 'comment posted' AND e.comment_length IS NOT NULL
+GROUP BY 1 ORDER BY 1`,
+				},
+				assert: (rows) => {
+					const by = cellsOf(rows, "bucket");
+					const sweet = by.sweet, low = by.low, over = by.over;
+					return guarded(
+						Number(sweet?.users) >= 400 && Number(low?.users) >= 80 && Number(over?.n_comments) >= 44000,
+						`buckets: sweet=${sweet?.users ?? 0}u low=${low?.users ?? 0}u over=${over?.n_comments ?? 0}c`,
+						() => {
+							const sweetOver = Number(sweet.len) / Number(over.len);
+							const sl = Number(sweet.len), ll = Number(low.len), ol = Number(over.len);
+							const detail = `comment_length sweet/over=${sweetOver.toFixed(3)} (measured 1.720); ordering sweet=${sl.toFixed(0)} low=${ll.toFixed(0)} over=${ol.toFixed(0)} (measured 250/200/145)`;
+							// Fix-round Q5 (S1): knob-implied sweet/over = POST_SWEET_COMMENT_BOOST(1.4)
+							// / POST_OVER_COMMENT_FACTOR(0.7) = 2.00; NAILED = knob ±10% = [1.80, 2.20].
+							// Final-count bucket migration (H2/H6/H8 clones push boosted sweet users into
+							// the 8+ bucket) attenuates the measured ratio to 1.72 — that reading grades
+							// STRONG by design. Ordering leg: NAILED only when both margins sit within
+							// knob ±10% (sweet/low ∈ [1.26, 1.54] around 1.40; low/over ∈ [1.29, 1.57]
+							// around 1/0.7 = 1.43); the 1.15-margin ordering grades STRONG.
+							const legRatio = bandVerdict(sweetOver, [1.80, 2.20], [1.40, 2.20], detail, v => v <= 1.05);
+							const mSweetLow = sl / ll, mLowOver = ll / ol;
+							let legOrder;
+							if (sl < ol) legOrder = { verdict: "INVERSE", detail };
+							else if (mSweetLow >= 1.26 && mSweetLow <= 1.54 && mLowOver >= 1.29 && mLowOver <= 1.57) legOrder = { verdict: "NAILED", detail };
+							else if (sl >= 1.15 * ll && ll >= 1.15 * ol) legOrder = { verdict: "STRONG", detail };
+							else legOrder = { verdict: "WEAK", detail };
+							return { verdict: worstOf(legRatio, legOrder), detail };
+						}
+					);
+				},
+			},
+		],
+	},
+	{
+		id: "social-h10-onboarding-ttc",
+		hook: "H10",
+		archetype: "funnel-ttc-by-segment",
+		narrative: "Creator/business complete onboarding faster than personal: knob-implied TTC ratio 0.568 (0.71/1.25); step-matching attenuation realizes ~0.71 — graded STRONG by design (window-stable 1h-24h).",
+		assertions: [
+			{
+				breakdown: {
+					type: "timeToConvert",
+					steps: ["account created", "profile updated", "post created"],
+					breakdownByUserProperty: "account_type",
+					conversionWindowMs: 6 * 3600 * 1000,
+				},
+				assert: (rows) => {
+					const by = cellsOf(rows, "segment_value");
+					const p = by.personal, c = by.creator, b = by.business;
+					const counts = `personal=${p?.user_count ?? 0} creator=${c?.user_count ?? 0} business=${b?.user_count ?? 0}`;
+					return guarded(
+						Number(p?.user_count) >= 160 && Number(c?.user_count) >= 125 && Number(b?.user_count) >= 115,
+						`converters: ${counts}`,
+						() => {
+							const fast = (Number(c.median_ttc_ms) + Number(b.median_ttc_ms)) / 2;
+							const ratio = fast / Number(p.median_ttc_ms);
+							const detail = `funnel median TTC @6h fast(creator/business)/personal=${ratio.toFixed(3)} (measured 0.709; medians ${(fast / 60000).toFixed(1)}min vs ${(Number(p.median_ttc_ms) / 60000).toFixed(1)}min; knobs 0.71/1.25 attenuate to 0.81/1.11 per leg)`;
+							// Fix-round Q5 (S1): knob-implied ratio = ONBOARDING_TTC_FAST(0.71) /
+							// ONBOARDING_TTC_SLOW(1.25) = 0.568; NAILED = knob ±10% = [0.52, 0.62].
+							// Greedy funnel step-matching substitutes organic standalone events and
+							// attenuates both legs toward 1 (0.81/1.11 realized — same mechanism as
+							// sass H9), by a non-knob-derivable amount; the measured 0.709 grades
+							// STRONG by design. Floor 0.51 = knob − 10% (mechanism can only attenuate
+							// toward 1, never overshoot); ceiling 0.85 sits clearly below the 0.963
+							// organic baseline.
+							return bandVerdict(ratio, [0.52, 0.62], [0.51, 0.85], detail, v => v >= 0.95);
+						}
+					);
+				},
+			},
+			{
+				breakdown: {
+					type: "duckdb",
+					sql: `SELECT COUNT(*) AS n,
+  AVG((user_id IS NOT NULL)::INT) AS uid_share,
+  AVG((device_id IS NOT NULL)::INT) AS device_share,
+  COUNT(DISTINCT device_id)::DOUBLE / COUNT(DISTINCT user_id) AS devices_per_user
+FROM ${EV}`,
+				},
+				assert: (rows) => {
+					const r = rows?.[0] || {};
+					return guarded(Number(r.n) >= 1300000, `events=${r.n ?? 0}`, () => {
+						const uid = Number(r.uid_share), dev = Number(r.device_share), dpu = Number(r.devices_per_user);
+						const detail = `identity invariants: uid_share=${uid} device_share=${dev.toFixed(4)} devices/user=${dpu.toFixed(2)} over ${r.n} events (auth on first event; avgDevicePerUser: 2)`;
+						// Fix-round Q5 (S1): dpu NAILED tightened to avgDevicePerUser(2) ±10%.
+						if (uid === 1 && dev >= 0.99 && dpu >= 1.8 && dpu <= 2.2) return { verdict: "NAILED", detail };
+						if (uid >= 0.999 && dev >= 0.98) return { verdict: "STRONG", detail };
+						if (uid < 0.9) return { verdict: "INVERSE", detail };
+						return { verdict: "WEAK", detail };
+					});
+				},
+			},
+		],
+	},
+];